@ Volodymyr Pavlyshyn
2024-12-07 14:52:47
The temporal semantics and **temporal and time-aware knowledge graphs. We have different memory models for artificial intelligence agents. We all try to mimic somehow how the brain works, or at least how the declarative memory of the brain works. We have the split of episodic memory** and **semantic memory**. And we also have a lot of theories, right?
## Declarative Memory of the Human Brain
How is the semantic memory formed? We all know that our brain stores semantic memory quite close to the concept we have with the personal knowledge graphs, that it’s connected entities. They form a connection with each other and all those things. So far, so good. And actually, then we have a lot of concepts, how the episodic memory and our experiences gets transmitted to the semantic:
- hippocampus indexing and retrieval
- sanitization of episodic memories
- episodic-semantic shift theory
They all give a different perspective on how different parts of declarative memory cooperate.
We know that episodic memories get semanticized over time. You have semantic knowledge without the notion of time, and probably, your episodic memory is just decayed.
But, you know, it’s still an open question:
> do we want to mimic an AI agent’s memory as a human brain memory, or do we want to create something different?
It’s an open question to which we have no good answer. And if you go to the theory of neuroscience and check how episodic and semantic memory interfere, you will still find a lot of theories, yeah?
Some of them say that you have the hippocampus that keeps the indexes of the memory. Some others will say that you semantic the episodic memory. Some others say that you have some separate process that digests the episodic and experience to the semantics. But all of them agree on the plan that it’s operationally two separate areas of memories and even two separate regions of brain, and the semantic, it’s more, let’s say, protected.
So it’s harder to forget the semantical facts than the episodes and everything. And what I’m thinking about for a long time, it’s this, you know, the semantic memory.
## Temporal Semantics
It’s memory about the facts, but you somehow mix the time information with the semantics. I already described a lot of things, including how we could combine time with knowledge graphs and how people do it.
There are multiple ways we could persist such information, but we all hit the wall because the complexity of time and the semantics of time are highly complex concepts.
## Time in a Semantic context is not a timestamp.
What I mean is that when you have a fact, and you just mentioned that I was there at this particular moment, like, I don’t know, 15:40 on Monday, it’s already awake because we don’t know which Monday, right? So you need to give the exact date, but usually, you do not have experiences like that.
You do not record your memories like that, except you do the journaling and all of the things. So, usually, you have no direct time references. What I mean is that you could say that I was there and it was some event, blah, blah, blah.
Somehow, we form a chain of events that connect with each other and maybe will be connected to some period of time if we are lucky enough. This means that we could not easily represent temporal-aware information as just a timestamp or validity and all of the things.
For sure, the validity of the knowledge graphs (simple quintuple with start and end dates)is a big topic, and it could solve a lot of things. It could solve a lot of the time cases. It’s super simple because you give the end and start dates, and you are done, but it does not answer facts that have a relative time or time information in facts . It could solve many use cases but struggle with facts in an indirect temporal context. I like the simplicity of this idea. But the problem of this approach that in most cases, we simply don’t have these timestamps. We don’t have the timestamp where this information starts and ends. And it’s not modeling many events in our life, especially if you have the processes or ongoing activities or recurrent events.
I’m more about thinking about the time of semantics, where you have a time model as a **hybrid clock** or some **global clock** that does the partial ordering of the events. It’s mean that you have the chain of the experiences and you have the chain of the facts that have the different time contexts.
We could deduct the time from this chain of the events. But it’s a big, big topic for the research. But what I want to achieve, actually, it’s not separation on episodic and semantic memory. It’s having something in between.
## Blockchain of connected events and facts
I call it temporal-aware semantics or time-aware knowledge graphs, where we could encode the semantic fact together with the time component.I doubt that time should be the simple timestamp or the region of the two timestamps. For me, it is more a chain for facts that have a partial order and form a blockchain like a database or a partially ordered Acyclic graph of facts that are temporally connected. We could have some notion of time that is understandable to the agent and a model that allows us to order the events and focus on what the agent knows and how to order this time knowledge and create the chains of the events.
## Time anchors
We may have a particular time in the chain that allows us to arrange a more concrete time for the rest of the events. But it’s still an open topic for research. The temporal semantics gets split into a couple of domains. One domain is how to add time to the knowledge graphs. We already have many different solutions. I described them in my previous articles.
Another domain is the agent's memory and how the memory of the artificial intelligence treats the time. This one, it’s much more complex. Because here, we could not operate with the simple timestamps. We need to have the representation of time that are understandable by model and understandable by the agent that will work with this model. And this one, it’s way bigger topic for the research.”